Research map / Architectures and efficiency / RNNs and normalization
Understanding the Disharmony Between Dropout and Batch Normalization by Variance Shift
Check your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply Dropout after all BN layers, using drop ratio 0.2 | ResNet-200, ResNeXt-101, and SENet · ImageNet validation | top-1 error: 21.70 -> 21.48; 20.40 -> 20.17; 18.89 -> 18.68 | Helped |
| Apply Dropout after all BN layers, using drop ratio 0.2 | ResNet-200, ResNeXt-101, and SENet · ImageNet validation | top-5 error: 5.80 -> 5.55; 5.30 -> 5.12; 4.66 -> 4.47 | Helped |
| Apply last-layer Dropout after all BN layers | PreResNet, ResNeXt, WRN, and DenseNet · CIFAR10 | error rate: 5.02 -> 4.94; 3.77 -> 3.69; 3.97 -> 3.84; 4.72 -> 4.67 | Mixed |
| Apply last-layer Dropout after all BN layers | PreResNet, ResNeXt, WRN, and DenseNet · CIFAR100 | error rate: 23.73 -> 23.43; 17.78 -> 17.77; 19.17 -> 19.17; 22.58 -> 21.86 | Mixed |
| Replace vanilla Dropout with Uout (ablation) | PreResNet, ResNeXt, WRN, and DenseNet in Dropout-(b) placement · CIFAR10 | error rate: 5.02 -> 4.85; 3.77 -> 3.75; 3.97 -> 3.80; 4.72 -> 4.61 | Mixed |
| Replace vanilla Dropout with Uout (ablation) | PreResNet, ResNeXt, WRN, and DenseNet in Dropout-(b) placement · CIFAR100 | error rate: 23.73 -> 23.53; 17.78 -> 17.74; 19.17 -> 18.87; 22.58 -> 22.30 | Mixed |
| Recalibrate BN moving statistics by running training data in evaluation mode (ablation) | Dropout-(a) and Dropout-(b) models with drop ratio 0.5 · CIFAR10 | error rate: PreResNet 8.42 -> 6.42; ResNeXt 4.43 -> 3.96; WRN 4.59 -> 4.20; DenseNet 8.70 -> 6.82 | Helped |
| Recalibrate BN moving statistics by running training data in evaluation mode (ablation) | Dropout-(a) and Dropout-(b) models with drop ratio 0.5 · CIFAR100 | error rate: PreResNet 32.45 -> 26.57; ResNeXt 19.04 -> 18.24; WRN 21.08 -> 20.70; DenseNet 31.45 -> 26.98 | Helped |
| Increase Dropout drop ratio to 0.5 in Dropout-(b) placement (ablation) | WRN · CIFAR10 | error rate | Helped |
| Use vanilla Dropout before BN in Dropout-(a) placement | DenseNet, PreResNet, ResNeXt, and WRN · CIFAR10 and CIFAR100 | error rate | Hurt |
Closest papers by meaning
- Dropout Reduces Underfitting (ICML 2023 · similarity 0.89)
- Understanding Batch Normalization (NeurIPS 2018 · similarity 0.88)
- The Implicit and Explicit Regularization Effects of Dropout (ICML 2020 · similarity 0.88)
- R-Drop: Regularized Dropout for Neural Networks (NeurIPS 2021 · similarity 0.88)
- On the Variance of Neural Network Training with respect to Test Sets and Distributions (ICLR 2024 · similarity 0.87)
- Stochastic Modified Equations and Dynamics of Dropout Algorithm (ICLR 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.